Heterogeneous Ensemble Models for In-Hospital Mortality Prediction

Abstract

Electronic Health Records data are rich and contain different types of variables, including structured data (e.g., demographics), free text (e.g., medical notes), and time series data. In this work, we explore the use of these different types of data for the task of in-hospital mortality prediction, which seeks to predict the outcome of death for patients admitted to the hospital. We build base learning models for the different data types and combine them in a heterogeneous ensemble model. In these models, we apply state-of-the-art classification algorithms based on deep learning. Our experiments on a set of 20K ICU patients from the MIMIC-III dataset showed that the ensemble method brings improvements of 3 percentage points, achieving an AUROC of 0.853.